He et al. (2026) Multi-scale feature fusion and uncertainty quantification in streamflow prediction: A temporal convolutional network approach with hybrid denoising

Identification

• Journal: Environmental Modelling & Software
• Year: 2026
• Date: 2026-01-19
• Authors: Yiping He, Z. Wang, Heqin Cheng, Weijie Ding
• DOI: 10.1016/j.envsoft.2026.106879

Research Groups

• College of Information, Shanghai Ocean University, Shanghai, China
• State Key Laboratory of Estuarine and Coastal Research, East China Normal University, Shanghai, China

Short Summary

This study proposes a novel hybrid deep learning model, NRBO-VMD-Wavelet-TCN (NVWT), for multi-scale streamflow prediction with uncertainty quantification. The model demonstrates superior short-term prediction accuracy (1–5 days) in the Hanjiang River Basin, outperforming benchmarks and providing interpretability.

Objective

• To develop a robust and interpretable hybrid deep learning model for accurate short-term streamflow prediction in complex watersheds, addressing challenges posed by nonlinearity, non-stationarity, and spatiotemporal heterogeneity, while also quantifying prediction uncertainty.

Study Configuration

• Spatial Scale: Nine hydrological stations within the Hanjiang River Basin, China.
• Temporal Scale: Short-term streamflow prediction (1–5 days ahead).

Methodology and Data

• Models used:
  • Proposed Model: NRBO-VMD-Wavelet-TCN (NVWT)
    • Newton-Raphson-based optimizer (NRBO) for adaptive Variational Mode Decomposition (VMD)
    • Wavelet Thresholding Denoising (WTD)
    • Temporal Convolutional Network (TCN) with dilated convolutions and residual connections
  • Feature Selection: Maximum Information Coefficient (MIC)
  • Interpretability: Shapley Additive Explanation (SHAP)
  • Benchmark Models: Gated Recurrent Unit (GRU), Transformer, Autoregressive Integrated Moving Average (ARIMA), Bidirectional Long Short-Term Memory (Bi-LSTM), Convolutional Neural Network (CNN), Ensemble Empirical Mode Decomposition (EEMD), Extreme Learning Machine (ELM), Empirical Mode Decomposition (EMD), Particle Swarm Optimization (PSO), Random Forests (RF).
• Data sources: Multi-source data, including upstream lagged streamflow and meteorological factors, used to construct a multidimensional feature system (lagged, periodic, and cumulative features).

Main Results

• The NVWT model achieved superior performance in short-term streamflow prediction (1–5 days) across all nine hydrological stations in the Hanjiang River Basin.
• Nash-Sutcliffe Efficiency (NSE) values exceeded 0.82 at all stations, peaking at 0.982 for the downstream Xiantao Station.
• The NVWT model significantly outperformed benchmark models such as GRU and Transformer.
• Ablation studies confirmed the effectiveness of the hybrid denoising module (NRBO-VMD and WTD) and the multidimensional feature system, particularly during flood peaks and low-flow periods.
• Interval prediction metrics validated the model's capability to quantify uncertainty in streamflow forecasts.
• SHAP analysis enhanced model interpretability by revealing the differential contributions of upstream lagged streamflow and meteorological factors to predictions.

Contributions

• Introduction of a novel hybrid deep learning model (NVWT) that integrates adaptive VMD with NRBO, Wavelet Thresholding Denoising, and TCN for enhanced streamflow prediction.
• Development of a robust multi-scale feature fusion strategy combined with MIC-based feature selection to improve input efficiency.
• Comprehensive uncertainty quantification through interval prediction metrics, providing more reliable forecasts.
• Enhanced model interpretability using SHAP, offering insights into the influence of various hydrological and meteorological factors.
• Provides a methodological reference for streamflow prediction in complex watersheds, with practical implications for flood control and water resource management under extreme climatic conditions.

Funding

• Not specified in the provided text.

Citation

@article{He2026Multiscale,
  author = {He, Yiping and Wang, Z. and Cheng, Heqin and Ding, Weijie},
  title = {Multi-scale feature fusion and uncertainty quantification in streamflow prediction: A temporal convolutional network approach with hybrid denoising},
  journal = {Environmental Modelling & Software},
  year = {2026},
  doi = {10.1016/j.envsoft.2026.106879},
  url = {https://doi.org/10.1016/j.envsoft.2026.106879}
}